1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright © 2019 Intel Corporation
4  */
5 
6 #include "i915_selftest.h"
7 #include "intel_engine_heartbeat.h"
8 #include "intel_engine_pm.h"
9 #include "intel_gt.h"
10 
11 #include "gem/selftests/mock_context.h"
12 #include "selftests/igt_flush_test.h"
13 #include "selftests/mock_drm.h"
14 
15 static int request_sync(struct i915_request *rq)
16 {
17 	struct intel_timeline *tl = i915_request_timeline(rq);
18 	long timeout;
19 	int err = 0;
20 
21 	intel_timeline_get(tl);
22 	i915_request_get(rq);
23 
24 	/* Opencode i915_request_add() so we can keep the timeline locked. */
25 	__i915_request_commit(rq);
26 	rq->sched.attr.priority = I915_PRIORITY_BARRIER;
27 	__i915_request_queue_bh(rq);
28 
29 	timeout = i915_request_wait(rq, 0, HZ / 10);
30 	if (timeout < 0)
31 		err = timeout;
32 	else
33 		i915_request_retire_upto(rq);
34 
35 	lockdep_unpin_lock(&tl->mutex, rq->cookie);
36 	mutex_unlock(&tl->mutex);
37 
38 	i915_request_put(rq);
39 	intel_timeline_put(tl);
40 
41 	return err;
42 }
43 
44 static int context_sync(struct intel_context *ce)
45 {
46 	struct intel_timeline *tl = ce->timeline;
47 	int err = 0;
48 
49 	mutex_lock(&tl->mutex);
50 	do {
51 		struct i915_request *rq;
52 		long timeout;
53 
54 		if (list_empty(&tl->requests))
55 			break;
56 
57 		rq = list_last_entry(&tl->requests, typeof(*rq), link);
58 		i915_request_get(rq);
59 
60 		timeout = i915_request_wait(rq, 0, HZ / 10);
61 		if (timeout < 0)
62 			err = timeout;
63 		else
64 			i915_request_retire_upto(rq);
65 
66 		i915_request_put(rq);
67 	} while (!err);
68 	mutex_unlock(&tl->mutex);
69 
70 	/* Wait for all barriers to complete (remote CPU) before we check */
71 	i915_active_unlock_wait(&ce->active);
72 	return err;
73 }
74 
75 static int __live_context_size(struct intel_engine_cs *engine)
76 {
77 	struct intel_context *ce;
78 	struct i915_request *rq;
79 	void *vaddr;
80 	int err;
81 
82 	ce = intel_context_create(engine);
83 	if (IS_ERR(ce))
84 		return PTR_ERR(ce);
85 
86 	err = intel_context_pin(ce);
87 	if (err)
88 		goto err;
89 
90 	vaddr = i915_gem_object_pin_map_unlocked(ce->state->obj,
91 						 i915_coherent_map_type(engine->i915,
92 									ce->state->obj, false));
93 	if (IS_ERR(vaddr)) {
94 		err = PTR_ERR(vaddr);
95 		intel_context_unpin(ce);
96 		goto err;
97 	}
98 
99 	/*
100 	 * Note that execlists also applies a redzone which it checks on
101 	 * context unpin when debugging. We are using the same location
102 	 * and same poison value so that our checks overlap. Despite the
103 	 * redundancy, we want to keep this little selftest so that we
104 	 * get coverage of any and all submission backends, and we can
105 	 * always extend this test to ensure we trick the HW into a
106 	 * compromising position wrt to the various sections that need
107 	 * to be written into the context state.
108 	 *
109 	 * TLDR; this overlaps with the execlists redzone.
110 	 */
111 	vaddr += engine->context_size - I915_GTT_PAGE_SIZE;
112 	memset(vaddr, POISON_INUSE, I915_GTT_PAGE_SIZE);
113 
114 	rq = intel_context_create_request(ce);
115 	intel_context_unpin(ce);
116 	if (IS_ERR(rq)) {
117 		err = PTR_ERR(rq);
118 		goto err_unpin;
119 	}
120 
121 	err = request_sync(rq);
122 	if (err)
123 		goto err_unpin;
124 
125 	/* Force the context switch */
126 	rq = intel_engine_create_kernel_request(engine);
127 	if (IS_ERR(rq)) {
128 		err = PTR_ERR(rq);
129 		goto err_unpin;
130 	}
131 	err = request_sync(rq);
132 	if (err)
133 		goto err_unpin;
134 
135 	if (memchr_inv(vaddr, POISON_INUSE, I915_GTT_PAGE_SIZE)) {
136 		pr_err("%s context overwrote trailing red-zone!", engine->name);
137 		err = -EINVAL;
138 	}
139 
140 err_unpin:
141 	i915_gem_object_unpin_map(ce->state->obj);
142 err:
143 	intel_context_put(ce);
144 	return err;
145 }
146 
147 static int live_context_size(void *arg)
148 {
149 	struct intel_gt *gt = arg;
150 	struct intel_engine_cs *engine;
151 	enum intel_engine_id id;
152 	int err = 0;
153 
154 	/*
155 	 * Check that our context sizes are correct by seeing if the
156 	 * HW tries to write past the end of one.
157 	 */
158 
159 	for_each_engine(engine, gt, id) {
160 		struct file *saved;
161 
162 		if (!engine->context_size)
163 			continue;
164 
165 		intel_engine_pm_get(engine);
166 
167 		/*
168 		 * Hide the old default state -- we lie about the context size
169 		 * and get confused when the default state is smaller than
170 		 * expected. For our do nothing request, inheriting the
171 		 * active state is sufficient, we are only checking that we
172 		 * don't use more than we planned.
173 		 */
174 		saved = fetch_and_zero(&engine->default_state);
175 
176 		/* Overlaps with the execlists redzone */
177 		engine->context_size += I915_GTT_PAGE_SIZE;
178 
179 		err = __live_context_size(engine);
180 
181 		engine->context_size -= I915_GTT_PAGE_SIZE;
182 
183 		engine->default_state = saved;
184 
185 		intel_engine_pm_put(engine);
186 
187 		if (err)
188 			break;
189 	}
190 
191 	return err;
192 }
193 
194 static int __live_active_context(struct intel_engine_cs *engine)
195 {
196 	unsigned long saved_heartbeat;
197 	struct intel_context *ce;
198 	int pass;
199 	int err;
200 
201 	/*
202 	 * We keep active contexts alive until after a subsequent context
203 	 * switch as the final write from the context-save will be after
204 	 * we retire the final request. We track when we unpin the context,
205 	 * under the presumption that the final pin is from the last request,
206 	 * and instead of immediately unpinning the context, we add a task
207 	 * to unpin the context from the next idle-barrier.
208 	 *
209 	 * This test makes sure that the context is kept alive until a
210 	 * subsequent idle-barrier (emitted when the engine wakeref hits 0
211 	 * with no more outstanding requests).
212 	 */
213 
214 	if (intel_engine_pm_is_awake(engine)) {
215 		pr_err("%s is awake before starting %s!\n",
216 		       engine->name, __func__);
217 		return -EINVAL;
218 	}
219 
220 	ce = intel_context_create(engine);
221 	if (IS_ERR(ce))
222 		return PTR_ERR(ce);
223 
224 	saved_heartbeat = engine->props.heartbeat_interval_ms;
225 	engine->props.heartbeat_interval_ms = 0;
226 
227 	for (pass = 0; pass <= 2; pass++) {
228 		struct i915_request *rq;
229 
230 		intel_engine_pm_get(engine);
231 
232 		rq = intel_context_create_request(ce);
233 		if (IS_ERR(rq)) {
234 			err = PTR_ERR(rq);
235 			goto out_engine;
236 		}
237 
238 		err = request_sync(rq);
239 		if (err)
240 			goto out_engine;
241 
242 		/* Context will be kept active until after an idle-barrier. */
243 		if (i915_active_is_idle(&ce->active)) {
244 			pr_err("context is not active; expected idle-barrier (%s pass %d)\n",
245 			       engine->name, pass);
246 			err = -EINVAL;
247 			goto out_engine;
248 		}
249 
250 		if (!intel_engine_pm_is_awake(engine)) {
251 			pr_err("%s is asleep before idle-barrier\n",
252 			       engine->name);
253 			err = -EINVAL;
254 			goto out_engine;
255 		}
256 
257 out_engine:
258 		intel_engine_pm_put(engine);
259 		if (err)
260 			goto err;
261 	}
262 
263 	/* Now make sure our idle-barriers are flushed */
264 	err = intel_engine_flush_barriers(engine);
265 	if (err)
266 		goto err;
267 
268 	/* Wait for the barrier and in the process wait for engine to park */
269 	err = context_sync(engine->kernel_context);
270 	if (err)
271 		goto err;
272 
273 	if (!i915_active_is_idle(&ce->active)) {
274 		pr_err("context is still active!");
275 		err = -EINVAL;
276 	}
277 
278 	intel_engine_pm_flush(engine);
279 
280 	if (intel_engine_pm_is_awake(engine)) {
281 		struct drm_printer p = drm_debug_printer(__func__);
282 
283 		intel_engine_dump(engine, &p,
284 				  "%s is still awake:%d after idle-barriers\n",
285 				  engine->name,
286 				  atomic_read(&engine->wakeref.count));
287 		GEM_TRACE_DUMP();
288 
289 		err = -EINVAL;
290 		goto err;
291 	}
292 
293 err:
294 	engine->props.heartbeat_interval_ms = saved_heartbeat;
295 	intel_context_put(ce);
296 	return err;
297 }
298 
299 static int live_active_context(void *arg)
300 {
301 	struct intel_gt *gt = arg;
302 	struct intel_engine_cs *engine;
303 	enum intel_engine_id id;
304 	int err = 0;
305 
306 	for_each_engine(engine, gt, id) {
307 		err = __live_active_context(engine);
308 		if (err)
309 			break;
310 
311 		err = igt_flush_test(gt->i915);
312 		if (err)
313 			break;
314 	}
315 
316 	return err;
317 }
318 
319 static int __remote_sync(struct intel_context *ce, struct intel_context *remote)
320 {
321 	struct i915_request *rq;
322 	int err;
323 
324 	err = intel_context_pin(remote);
325 	if (err)
326 		return err;
327 
328 	rq = intel_context_create_request(ce);
329 	if (IS_ERR(rq)) {
330 		err = PTR_ERR(rq);
331 		goto unpin;
332 	}
333 
334 	err = intel_context_prepare_remote_request(remote, rq);
335 	if (err) {
336 		i915_request_add(rq);
337 		goto unpin;
338 	}
339 
340 	err = request_sync(rq);
341 
342 unpin:
343 	intel_context_unpin(remote);
344 	return err;
345 }
346 
347 static int __live_remote_context(struct intel_engine_cs *engine)
348 {
349 	struct intel_context *local, *remote;
350 	unsigned long saved_heartbeat;
351 	int pass;
352 	int err;
353 
354 	/*
355 	 * Check that our idle barriers do not interfere with normal
356 	 * activity tracking. In particular, check that operating
357 	 * on the context image remotely (intel_context_prepare_remote_request),
358 	 * which inserts foreign fences into intel_context.active, does not
359 	 * clobber the idle-barrier.
360 	 */
361 
362 	if (intel_engine_pm_is_awake(engine)) {
363 		pr_err("%s is awake before starting %s!\n",
364 		       engine->name, __func__);
365 		return -EINVAL;
366 	}
367 
368 	remote = intel_context_create(engine);
369 	if (IS_ERR(remote))
370 		return PTR_ERR(remote);
371 
372 	local = intel_context_create(engine);
373 	if (IS_ERR(local)) {
374 		err = PTR_ERR(local);
375 		goto err_remote;
376 	}
377 
378 	saved_heartbeat = engine->props.heartbeat_interval_ms;
379 	engine->props.heartbeat_interval_ms = 0;
380 	intel_engine_pm_get(engine);
381 
382 	for (pass = 0; pass <= 2; pass++) {
383 		err = __remote_sync(local, remote);
384 		if (err)
385 			break;
386 
387 		err = __remote_sync(engine->kernel_context, remote);
388 		if (err)
389 			break;
390 
391 		if (i915_active_is_idle(&remote->active)) {
392 			pr_err("remote context is not active; expected idle-barrier (%s pass %d)\n",
393 			       engine->name, pass);
394 			err = -EINVAL;
395 			break;
396 		}
397 	}
398 
399 	intel_engine_pm_put(engine);
400 	engine->props.heartbeat_interval_ms = saved_heartbeat;
401 
402 	intel_context_put(local);
403 err_remote:
404 	intel_context_put(remote);
405 	return err;
406 }
407 
408 static int live_remote_context(void *arg)
409 {
410 	struct intel_gt *gt = arg;
411 	struct intel_engine_cs *engine;
412 	enum intel_engine_id id;
413 	int err = 0;
414 
415 	for_each_engine(engine, gt, id) {
416 		err = __live_remote_context(engine);
417 		if (err)
418 			break;
419 
420 		err = igt_flush_test(gt->i915);
421 		if (err)
422 			break;
423 	}
424 
425 	return err;
426 }
427 
428 int intel_context_live_selftests(struct drm_i915_private *i915)
429 {
430 	static const struct i915_subtest tests[] = {
431 		SUBTEST(live_context_size),
432 		SUBTEST(live_active_context),
433 		SUBTEST(live_remote_context),
434 	};
435 	struct intel_gt *gt = &i915->gt;
436 
437 	if (intel_gt_is_wedged(gt))
438 		return 0;
439 
440 	return intel_gt_live_subtests(tests, gt);
441 }
442